Diagnostic and Prognostic Analysis Optimization of Field Problems for EV Charging Stations¶

X. Zhang — M.Sc. thesis, Aalto University, August 2020 (with ABB)

A DC fast charger is a distributed system of power cabinets, charger posts and five separate control boards communicating over a controller area network (CAN) bus. When one fails, a service engineer has to work out which part. Engineers can only reach a few chargers a day, so every hour spent diagnosing is an hour the charger is down. This thesis asks how much of that diagnosis can be done from sensor data the charger already emits.

The learning problem¶

A large share of the work here is upstream of any model, and the thesis is structured around that: identify which sensor data reflects charger condition, get it off the device, and turn a CAN bus trace into a dataset. Only then is there a learning problem at all.

Getting the data. Messages are captured from the CAN bus with a logger. The bus offers two mechanisms, and the choice between them matters: Service Data Objects allow access to only one dictionary object at a time, which is prohibitive for continuously changing values, so Process Data Objects are used instead. Each PDO message carries 8 bytes per node per timestamp, parsed against ABB's design documentation.

Data points. One data point is a single charging session of one charger.

Features. Five, extracted from the CAN trace: charging current, charging voltage, fault resistance, board humidity and board temperature. The per-feature normalisation is concrete: charging current is divided by 150 A, the expected operating range; voltage by its own maximum, since it stabilises quickly; fault resistance is measured as deviation from its expected 2500 kΩ. Each is summarised by its maximum, minimum and average over the session, and normalised to [0, 1]. The normalisation method is chosen per feature rather than globally, so that each signal keeps its own minimum, maximum and spread, and the pipeline does not depend on specific hardware or software.

Labels. There are none at training time. The thesis works through the supervised, semi-supervised and unsupervised anomaly-detection setups and argues for the last: a supervised detector only finds faults whose signature is known in advance, which is the wrong assumption for a complex charger where novel faults appear. Labels derived from each session's stop reason are used only to evaluate the result.

Hypothesis space. DBSCAN — density-based spatial clustering of applications with noise — over the five normalised features. Points falling outside any dense region are the suspected faults. Performance was best at ε = 0.10 and MinPts = 4, chosen over five experiments fine-tuning those two parameters.

Objective. No loss is minimised: DBSCAN fits nothing. Results are scored against the stop-reason labels on five metrics, precision and recall among them.

What it found¶

The thesis frames its result as an answer to "to what extent can machine learning help", deliberately a question of degree. Its main contribution is the pipeline rather than the classifier: a reproducible path from raw CAN bus traffic on deployed hardware to a normalised, labelled dataset, including the engineering judgement about which bus protocol makes continuous monitoring feasible. That part transfers to any CAN-based industrial system; the specific classifier does not.

Watch the defence¶

Full text¶

The thesis record is at Aaltodoc. The full text is not open access.